Dynamic fault diagnosis using extended matrix and tensor locality preserving discriminant analysis

It is well acknowledged that utilizing dynamic information can improve accuracy in fault diagnosis for dynamic processes. Conventional methods encode dynamic information by constructing an extended vector comprising current process data as well as past process data. Then the classic Linear Discriminant Analysis (LDA) is usually applied to extended vectors to reduce dimensionality and obtain a discriminative subspace where overlapping among different fault classes is minimized. However, using extended vectors aggravates the “curse of dimensionality” problem and loses structure information in variables. Besides, LDA probably provides suboptimal results when there are more than two candidate fault classes. This paper proposes using extended matrices to encoding dynamic information and using a novel dimensionality reduction method named Tensor Locality Preserving Discriminant Analysis (TLPDA) to perform dimensionality reduction on extend matrices directly. TLPDA is based on local structure in data and overcomes the main drawbacks of LDA. A new dynamic fault diagnosis scheme is developed based on extended matrices and TLPDA. Extensive simulations on the Tennessee Eastman (TE) benchmark simulation process clearly demonstrate the superiority of our methods in terms of misclassification rate and making use of extra training data.